Source code for evennia.utils.idmapper.models

Django ID mapper

Modified for Evennia by making sure that no model references
leave caching unexpectedly (no use of WeakRefs).

Also adds `cache_size()` for monitoring the size of the cache.

import gc
import os
import threading
import time
from weakref import WeakValueDictionary

from django.core.exceptions import FieldError, ObjectDoesNotExist
from django.db.models.base import Model, ModelBase
from django.db.models.signals import post_migrate, post_save, pre_delete
from django.db.transaction import atomic
from django.db.utils import DatabaseError
from twisted.internet.reactor import callFromThread

from evennia.utils import logger
from evennia.utils.utils import dbref, get_evennia_pids, to_str

from .manager import SharedMemoryManager

AUTO_FLUSH_MIN_INTERVAL = 60.0 * 5  # at least 5 mins between cache flushes

_GA = object.__getattribute__
_SA = object.__setattr__
_DA = object.__delattr__

# References to db-updated objects are stored here so the
# main process can be informed to re-cache itself.
PROC_MODIFIED_OBJS = WeakValueDictionary()

# get info about the current process and thread; determine if our
# current pid is different from the server PID (i.e.  # if we are in a
# subprocess or not)
_SELF_PID = os.getpid()
_SERVER_PID, _PORTAL_PID = get_evennia_pids()
_IS_MAIN_THREAD = threading.current_thread().name == "MainThread"

[docs]class SharedMemoryModelBase(ModelBase): # CL: upstream had a __new__ method that skipped ModelBase's __new__ if # SharedMemoryModelBase was not in the model class's ancestors. It's not # clear what was the intended purpose, but skipping ModelBase.__new__ # broke things; in particular, default manager inheritance. def __call__(cls, *args, **kwargs): """ this method will either create an instance (by calling the default implementation) or try to retrieve one from the class-wide cache by inferring the pk value from `args` and `kwargs`. If instance caching is enabled for this class, the cache is populated whenever possible (ie when it is possible to infer the pk value). """ def new_instance(): return super(SharedMemoryModelBase, cls).__call__(*args, **kwargs) instance_key = cls._get_cache_key(args, kwargs) # depending on the arguments, we might not be able to infer the PK, so in that case we # create a new instance if instance_key is None: return new_instance() cached_instance = cls.get_cached_instance(instance_key) if cached_instance is None: cached_instance = new_instance() cls.cache_instance(cached_instance, new=True) return cached_instance def _prepare(cls): """ Prepare the cache, making sure that proxies of the same db base share the same cache. """ # the dbmodel is either the proxy base or ourselves dbmodel = cls._meta.concrete_model if cls._meta.proxy else cls cls.__dbclass__ = dbmodel if not hasattr(dbmodel, "__instance_cache__"): # we store __instance_cache__ only on the dbmodel base dbmodel.__instance_cache__ = {} super()._prepare() def __new__(cls, name, bases, attrs): """ Field shortcut creation: Takes field names `db_*` and creates property wrappers named without the `db_` prefix. So db_key -> key This wrapper happens on the class level, so there is no overhead when creating objects. If a class already has a wrapper of the given name, the automatic creation is skipped. Notes: Remember to document this auto-wrapping in the class header, this could seem very much like magic to the user otherwise. """ attrs["typename"] = cls.__name__ attrs["path"] = "%s.%s" % (attrs["__module__"], name) attrs["_is_deleted"] = False # set up the typeclass handling only if a variable _is_typeclass is set on the class def create_wrapper(cls, fieldname, wrappername, editable=True, foreignkey=False): "Helper method to create property wrappers with unique names (must be in separate call)" def _get(cls, fname): "Wrapper for getting database field" if _GA(cls, "_is_deleted"): raise ObjectDoesNotExist( "Cannot access %s: Hosting object was already deleted." % fname ) return _GA(cls, fieldname) def _get_foreign(cls, fname): "Wrapper for returning foreignkey fields" if _GA(cls, "_is_deleted"): raise ObjectDoesNotExist( "Cannot access %s: Hosting object was already deleted." % fname ) return _GA(cls, fieldname) def _set_nonedit(cls, fname, value): "Wrapper for blocking editing of field" raise FieldError("Field %s cannot be edited." % fname) def _set(cls, fname, value): "Wrapper for setting database field" if _GA(cls, "_is_deleted"): raise ObjectDoesNotExist( "Cannot set %s to %s: Hosting object was already deleted!" % (fname, value) ) _SA(cls, fname, value) # only use explicit update_fields in save if we actually have a # primary key assigned already (won't be set when first creating object) update_fields = ( [fname] if _GA(cls, "_get_pk_val")(_GA(cls, "_meta")) is not None else None ) _GA(cls, "save")(update_fields=update_fields) def _set_foreign(cls, fname, value): "Setter only used on foreign key relations, allows setting with #dbref" if _GA(cls, "_is_deleted"): raise ObjectDoesNotExist( "Cannot set %s to %s: Hosting object was already deleted!" % (fname, value) ) if isinstance(value, (str, int)): value = to_str(value) if value.isdigit() or value.startswith("#"): # we also allow setting using dbrefs, if so we try to load the matching # object. (we assume the object is of the same type as the class holding # the field, if not a custom handler must be used for that field) dbid = dbref(value, reqhash=False) if dbid: model = _GA(cls, "_meta").get_field(fname).model try: value = model._default_manager.get(id=dbid) except ObjectDoesNotExist: # maybe it is just a name that happens to look like a dbid pass _SA(cls, fname, value) # only use explicit update_fields in save if we actually have a # primary key assigned already (won't be set when first creating object) update_fields = ( [fname] if _GA(cls, "_get_pk_val")(_GA(cls, "_meta")) is not None else None ) _GA(cls, "save")(update_fields=update_fields) def _del_nonedit(cls, fname): "wrapper for not allowing deletion" raise FieldError("Field %s cannot be edited." % fname) def _del(cls, fname): "Wrapper for clearing database field - sets it to None" _SA(cls, fname, None) update_fields = ( [fname] if _GA(cls, "_get_pk_val")(_GA(cls, "_meta")) is not None else None ) _GA(cls, "save")(update_fields=update_fields) # wrapper factories if not editable: def fget(cls): return _get(cls, fieldname) def fset(cls, val): return _set_nonedit(cls, fieldname, val) elif foreignkey: def fget(cls): return _get_foreign(cls, fieldname) def fset(cls, val): return _set_foreign(cls, fieldname, val) else: def fget(cls): return _get(cls, fieldname) def fset(cls, val): return _set(cls, fieldname, val) def fdel(cls): return _del(cls, fieldname) if editable else _del_nonedit(cls, fieldname) # set docstrings for auto-doc fget.__doc__ = "A wrapper for getting database field `%s`." % fieldname fset.__doc__ = "A wrapper for setting (and saving) database field `%s`." % fieldname fdel.__doc__ = "A wrapper for deleting database field `%s`." % fieldname # assigning attrs[wrappername] = property(fget, fset, fdel) # type(cls).__setattr__(cls, wrappername, property(fget, fset, fdel))#, doc)) # exclude some models that should not auto-create wrapper fields if cls.__name__ in ("ServerConfig", "TypeNick"): return # dynamically create the wrapper properties for all fields not already handled # (manytomanyfields are always handlers) for fieldname, field in ( (fname, field) for fname, field in list(attrs.items()) if fname.startswith("db_") and type(field).__name__ != "ManyToManyField" ): foreignkey = type(field).__name__ == "ForeignKey" wrappername = "dbid" if fieldname == "id" else fieldname.replace("db_", "", 1) if wrappername not in attrs: # makes sure not to overload manually created wrappers on the model create_wrapper( cls, fieldname, wrappername, editable=field.editable, foreignkey=foreignkey ) return super().__new__(cls, name, bases, attrs)
[docs]class SharedMemoryModel(Model, metaclass=SharedMemoryModelBase): """ Base class for idmapped objects. Inherit from `this`. """ objects = SharedMemoryManager()
[docs] class Meta(object): abstract = True
@classmethod def _get_cache_key(cls, args, kwargs): """ This method is used by the caching subsystem to infer the PK value from the constructor arguments. It is used to decide if an instance has to be built or is already in the cache. """ result = None # Quick hack for my composites work for now. if hasattr(cls._meta, "pks"): pk = cls._meta.pks[0] else: pk = # get the index of the pk in the class fields. this should be calculated *once*, but isn't # atm pk_position = cls._meta.fields.index(pk) if len(args) > pk_position: # if it's in the args, we can get it easily by index result = args[pk_position] elif pk.attname in kwargs: # retrieve the pk value. Note that we use attname instead of name, to handle the case # where the pk is a a ForeignKey. result = kwargs[pk.attname] elif != pk.attname and in kwargs: # ok we couldn't find the value, but maybe it's a FK and we can find the corresponding # object instead result = kwargs[] if result is not None and isinstance(result, Model): # if the pk value happens to be a model instance (which can happen wich a FK), we'd # rather use its own pk as the key result = result._get_pk_val() return result
[docs] @classmethod def get_cached_instance(cls, id): """ Method to retrieve a cached instance by pk value. Returns None when not found (which will always be the case when caching is disabled for this class). Please note that the lookup will be done even when instance caching is disabled. """ return cls.__dbclass__.__instance_cache__.get(id)
[docs] @classmethod def cache_instance(cls, instance, new=False): """ Method to store an instance in the cache. Args: instance (Class instance): the instance to cache. new (bool, optional): this is the first time this instance is cached (i.e. this is not an update operation like after a db save). """ pk = instance._get_pk_val() if pk is not None: new = new or pk not in cls.__dbclass__.__instance_cache__ cls.__dbclass__.__instance_cache__[pk] = instance if new: try: # trigger the at_init hook only # at first initialization instance.at_init() except AttributeError: # The at_init hook is not assigned to all entities pass
[docs] @classmethod def get_all_cached_instances(cls): """ Return the objects so far cached by idmapper for this class. """ return list(cls.__dbclass__.__instance_cache__.values())
@classmethod def _flush_cached_by_key(cls, key, force=True): """ Remove the cached reference. """ try: if force or cls.at_idmapper_flush(): del cls.__dbclass__.__instance_cache__[key] else: cls._dbclass__.__instance_cache__[key].refresh_from_db() except KeyError: # No need to remove if cache doesn't contain it already pass
[docs] @classmethod def flush_cached_instance(cls, instance, force=True): """ Method to flush an instance from the cache. The instance will always be flushed from the cache, since this is most likely called from delete(), and we want to make sure we don't cache dead objects. """ cls._flush_cached_by_key(instance._get_pk_val(), force=force)
# flush_cached_instance = classmethod(flush_cached_instance)
[docs] @classmethod def flush_instance_cache(cls, force=False): """ This will clean safe objects from the cache. Use `force` keyword to remove all objects, safe or not. """ if force: cls.__dbclass__.__instance_cache__ = {} else: cls.__dbclass__.__instance_cache__ = dict( (key, obj) for key, obj in cls.__dbclass__.__instance_cache__.items() if not obj.at_idmapper_flush() )
# flush_instance_cache = classmethod(flush_instance_cache) # per-instance methods def __eq__(self, other): return super().__eq__(other) def __hash__(self): # this is required to maintain hashing return super().__hash__()
[docs] def at_idmapper_flush(self): """ This is called when the idmapper cache is flushed and allows customized actions when this happens. Returns: do_flush (bool): If True, flush this object as normal. If False, don't flush and expect this object to handle the flushing on its own. """ return True
[docs] def flush_from_cache(self, force=False): """ Flush this instance from the instance cache. Use `force` to override the result of at_idmapper_flush() for the object. """ pk = self._get_pk_val() if pk: if force or self.at_idmapper_flush(): self.__class__.__dbclass__.__instance_cache__.pop(pk, None)
[docs] def delete(self, *args, **kwargs): """ Delete the object, clearing cache. """ self.flush_from_cache() self._is_deleted = True super().delete(*args, **kwargs)
[docs] def save(self, *args, **kwargs): """ Central database save operation. Notes: Arguments as per Django documentation. Calls `self.at_<fieldname>_postsave(new)` (this is a wrapper set by oobhandler: self._oob_at_<fieldname>_postsave()) """ global _MONITOR_HANDLER if not _MONITOR_HANDLER: from evennia.scripts.monitorhandler import ( MONITOR_HANDLER as _MONITOR_HANDLER, ) if _IS_SUBPROCESS: # we keep a store of objects modified in subprocesses so # we know to update their caches in the central process global PROC_MODIFIED_COUNT, PROC_MODIFIED_OBJS PROC_MODIFIED_COUNT += 1 PROC_MODIFIED_OBJS[PROC_MODIFIED_COUNT] = self if _IS_MAIN_THREAD: # in main thread - normal operation try: with atomic(): super().save(*args, **kwargs) except DatabaseError: # we handle the 'update_fields did not update any rows' error that # may happen due to timing issues with attributes ufields_removed = kwargs.pop("update_fields", None) if ufields_removed: with atomic(): super().save(*args, **kwargs) else: raise else: # in another thread; make sure to save in reactor thread def _save_callback(cls, *args, **kwargs): super().save(*args, **kwargs) callFromThread(_save_callback, self, *args, **kwargs) if not # this can happen if some of the startup methods immediately # delete the object (an example are Scripts that start and die immediately) return # update field-update hooks and eventual OOB watchers new = False if "update_fields" in kwargs and kwargs["update_fields"]: # get field objects from their names update_fields = ( self._meta.get_field(fieldname) for fieldname in kwargs.get("update_fields") ) else: # meta.fields are already field objects; get them all new = True update_fields = self._meta.fields for field in update_fields: fieldname = # trigger eventual monitors _MONITOR_HANDLER.at_update(self, fieldname) # if a hook is defined it must be named exactly on this form hookname = "at_%s_postsave" % fieldname if hasattr(self, hookname) and callable(_GA(self, hookname)): _GA(self, hookname)(new) # # if a trackerhandler is set on this object, update it with the # # fieldname and the new value # fieldtracker = "_oob_at_%s_postsave" % fieldname # if hasattr(self, fieldtracker): # _GA(self, fieldtracker)(fieldname) pass
[docs]class WeakSharedMemoryModelBase(SharedMemoryModelBase): """ Uses a WeakValue dictionary for caching instead of a regular one. """ def _prepare(cls): super()._prepare() cls.__dbclass__.__instance_cache__ = WeakValueDictionary()
[docs]class WeakSharedMemoryModel(SharedMemoryModel, metaclass=WeakSharedMemoryModelBase): """ Uses a WeakValue dictionary for caching instead of a regular one """
[docs] class Meta(object): abstract = True
[docs]def flush_cache(**kwargs): """ Flush idmapper cache. When doing so the cache will fire the at_idmapper_flush hook to allow the object to optionally handle its own flushing. Uses a signal so we make sure to catch cascades. """ def class_hierarchy(clslist): """Recursively yield a class hierarchy""" for cls in clslist: subclass_list = cls.__subclasses__() if subclass_list: for subcls in class_hierarchy(subclass_list): yield subcls else: yield cls for cls in class_hierarchy([SharedMemoryModel]): cls.flush_instance_cache() # run the python garbage collector return gc.collect()
# request_finished.connect(flush_cache) post_migrate.connect(flush_cache)
[docs]def flush_cached_instance(sender, instance, **kwargs): """ Flush the idmapper cache only for a given instance. """ # XXX: Is this the best way to make sure we can flush? if not hasattr(instance, "flush_cached_instance"): return sender.flush_cached_instance(instance, force=True)
[docs]def update_cached_instance(sender, instance, **kwargs): """ Re-cache the given instance in the idmapper cache. """ if not hasattr(instance, "cache_instance"): return sender.cache_instance(instance)
post_save.connect(update_cached_instance) LAST_FLUSH = None
[docs]def conditional_flush(max_rmem, force=False): """ Flush the cache if the estimated memory usage exceeds `max_rmem`. The flusher has a timeout to avoid flushing over and over in particular situations (this means that for some setups the memory usage will exceed the requirement and a server with more memory is probably required for the given game). Args: max_rmem (int): memory-usage estimation-treshold after which cache is flushed. force (bool, optional): forces a flush, regardless of timeout. Defaults to `False`. """ global LAST_FLUSH def mem2cachesize(desired_rmem): """ Estimate the size of the idmapper cache based on the memory desired. This is used to optionally cap the cache size. desired_rmem - memory in MB (minimum 50MB) The formula is empirically estimated from usage tests (Linux) and is Ncache = RMEM - 35.0 / 0.0157 where RMEM is given in MB and Ncache is the size of the cache for this memory usage. VMEM tends to be about 100MB higher than RMEM for large memory usage. """ vmem = max(desired_rmem, 50.0) Ncache = int(abs(float(vmem) - 35.0) / 0.0157) return Ncache if not max_rmem: # auto-flush is disabled return now = time.time() if not LAST_FLUSH: # server is just starting LAST_FLUSH = now return if ((now - LAST_FLUSH) < AUTO_FLUSH_MIN_INTERVAL) and not force: # too soon after last flush. logger.log_warn( "Warning: Idmapper flush called more than once in %s min interval. Check memory usage." % (AUTO_FLUSH_MIN_INTERVAL / 60.0) ) return if == "nt": # we can't look for mem info in Windows at the moment return # check actual memory usage Ncache_max = mem2cachesize(max_rmem) Ncache, _ = cache_size() actual_rmem = ( float(os.popen("ps -p %d -o %s | tail -1" % (os.getpid(), "rss")).read()) / 1000.0 ) # resident memory if Ncache >= Ncache_max and actual_rmem > max_rmem * 0.9: # flush cache when number of objects in cache is big enough and our # actual memory use is within 10% of our set max flush_cache() LAST_FLUSH = now
[docs]def cache_size(mb=True): """ Calculate statistics about the cache. Note: we cannot get reliable memory statistics from the cache - whereas we could do `getsizof` each object in cache, the result is highly imprecise and for a large number of objects the result is many times larger than the actual memory usage of the entire server; Python is clearly reusing memory behind the scenes that we cannot catch in an easy way here. Ideas are appreciated. /Griatch Returns: total_num, {objclass:total_num, ...} """ numtotal = [0] # use mutable to keep reference through recursion classdict = {} def get_recurse(submodels): for submodel in submodels: subclasses = submodel.__subclasses__() if not subclasses: num = len(submodel.get_all_cached_instances()) numtotal[0] += num classdict[submodel.__dbclass__.__name__] = num else: get_recurse(subclasses) get_recurse(SharedMemoryModel.__subclasses__()) return numtotal[0], classdict